If your goal is that each call to rand(), regardless of process,
sequentially pulls a number from the RNG stream, then perhaps you just need
to create a MersenneTwister([seed]) object.
mystream = MersenneTwister(1)
## Parallel execution
srand(1)
parfor = @parallel (vcat) for i=1:4
rand(mystream)
end
Intuitively, this seems like it must come with some communication overhead.
It seems likely that there should be a way to avoid this overhead with
pmap().
Jim
On Tuesday, July 15, 2014 6:25:11 PM UTC-4, Gray Calhoun wrote:
>
> Hi everyone, I'm trying to start using Julia for some Monte Carlo
> simulations
> (not MCMC) which I'd like to parallelize. I haven't found any documentation
> for setting the RNG's seed for parallelization. The naive approach gives
> different results than non-parallel execution (which is not surprising).
>
> Starting with `julia -p 4` and executing:
>
> ## Sequential execution
> srand(1)
> seqfor = Array(Float64,4)
> for i=1:4
> seqfor[i] = rand()
> end
>
> ## Parallel execution
> srand(1)
> parfor = @parallel (vcat) for i=1:4
> rand()
> end
>
> [sort(parfor) sort(seqfor)]
>
> gives
>
> 4x2 Array{Float64,2}:
> 0.346517 0.00790928
> 0.346517 0.236033
> 0.662369 0.312707
> 0.914194 0.346517
>
> and re-running the parallel code can give different results even after
> re-seeding. If we start julia without `-p 4` then both loops give the
> same results. If it matters, I'm using Julia version 0.3.0-rc1+28
> from source (commit 79e4771).
>
> Is there documentation on the right way to parallelize simulations in
> Julia? If not, should my next step be to carefully read the "parallel
> computing" documentation?
>
